Match BOX idioms early in the importer

[jkotas]

This avoids some of the surprise allocations mentioned in #22984, and helps with downstream optimizations in some cases.

[jkotas]

Number of small improvements in CoreLib R2R image (the CoreLib image is about 2kB smaller with this change). For example:

; Assembly listing for method ThrowHelper:IfNullAndNullsAreIllegalThenThrow(ref,int)

Before:

       sub      rsp, 40
       mov      qword ptr [rsp+20H], rcx

G_M57383_IG02:
       test     rdx, rdx
       je       SHORT G_M57383_IG04

G_M57383_IG03:
       add      rsp, 40
       ret      

G_M57383_IG04:
       mov      ecx, r8d
       call     [ThrowHelper:ThrowArgumentNullException(int)]
       int3     

After:

G_M57383_IG01:
       sub      rsp, 40
       nop      // Argument is not spilled to stack anymore

G_M57383_IG02:
       test     rdx, rdx
       je       SHORT G_M57383_IG04

G_M57383_IG03:
       add      rsp, 40
       ret      

G_M57383_IG04:
       mov      ecx, r8d
       call     [ThrowHelper:ThrowArgumentNullException(int)]
       int3     

[jkotas]

@AndyAyersMS PTLA

[jkotas]

Would it be better to just spill the side-effects here instead?

[AndyAyersMS]

I think it's cleaner like you have it -- leaves the eval stack unmodified when the patterns don't match.

[mikedn]

Yay, the JIT team got bigger 😄

[MarcoRossignoli]

Someone should write a "detailed "book on it, there are a lot of info/book on GC but few info(other than code and old high level BOTR) of this great piece of tech.

[AndyAyersMS]

Changes LGTM.

Interesting that this led to diffs even when optimizing. I would have expected we'd end up with teh same code. Did you look into why this happens? If not, I'll put it on my todo list.

[jkotas]

Did you look into why this happens?

I have looked into a few of them:

• The box instruction can fail inlining in shared generic code because of dictionary lookup. When we pick up the box idiom early, the inlining will succeed. IfNullAndNullsAreIllegalThenThrow is method that hits it.
• The extra locals created by the box expansion do not disappear and affect register allocation.